Publish provider benchmark results: 1k to 1M scales

- results-gemini-1m: Amazon Polarity 1M/7k, provider gemini 3.7-flash with
  live 20-row sampled audit (8.98/10) and 20/20 classification; recovery
  +0.065, clean delta +0.000
- results-openai-25k/50k/100k: AG News ladder (+0.038/+0.028/+0.028) and
  imdb +0.109 at 25k, provider openai gpt-5.6-sol config, local stages
- results-claude/results-openai: 1k baselines for the scale comparison

Author: voidstackloop

benchmarks/results-claude/REPORT.md

@@ -0,0 +1,18 @@
+# Claude benchmark (Anthropic)
+
+Provider: `anthropic` | Model: `claude-sonnet-5` | Base URL: `-` | Seeds: [17, 29] | Epochs: 2 |
+Audit rows: 0 | Classify sample: 20
+
+## Recovery (Claude-audited dirty_optimized vs dirty_raw)
+
+| Dataset | Classes | Dirty raw | Dirty optimized | Recovery | Clean delta |
+|---|---|---:|---:|---:|---:|
+| ag_news | 4 | 0.6590 | 0.6180 | -0.0410 | +0.0000 |
+| imdb | 2 | 0.5920 | 0.6850 | +0.0930 | +0.0000 |
+
+## Claude usage per dataset
+
+| Dataset | Audit tokens (in/out/total) | Audit score | Classify tokens |
+|---|---|---|---|
+| ag_news | 0/0/0 | None | 0 |
+| imdb | 0/0/0 | None | 0 |